Emulate double precision for regular rendering operation.

We calculate the lost precision on the CPU and pass it into the GPU
so that it can calculate an error-corrected version of the vertex position

1 files changed, 77 insertions, 4 deletions

diff --git a/servers/rendering/renderer_rd/shaders/forward_clustered/scene_forward_clustered.glsl b/servers/rendering/renderer_rd/shaders/forward_clustered/scene_forward_clustered.glsl
index d41474118d..adc4d9d01e 100644
--- a/servers/rendering/renderer_rd/shaders/forward_clustered/scene_forward_clustered.glsl
+++ b/servers/rendering/renderer_rd/shaders/forward_clustered/scene_forward_clustered.glsl
@@ -129,12 +129,52 @@ invariant gl_Position;
 
 #GLOBALS
 
+#ifdef USE_DOUBLE_PRECISION
+// Helper functions for emulating double precision when adding floats.
+vec3 quick_two_sum(vec3 a, vec3 b, out vec3 out_p) {
+	vec3 s = a + b;
+	out_p = b - (s - a);
+	return s;
+}
+
+vec3 two_sum(vec3 a, vec3 b, out vec3 out_p) {
+	vec3 s = a + b;
+	vec3 v = s - a;
+	out_p = (a - (s - v)) + (b - v);
+	return s;
+}
+
+vec3 double_add_vec3(vec3 base_a, vec3 prec_a, vec3 base_b, vec3 prec_b, out vec3 out_precision) {
+	vec3 s, t, se, te;
+	s = two_sum(base_a, base_b, se);
+	t = two_sum(prec_a, prec_b, te);
+	se += t;
+	s = quick_two_sum(s, se, se);
+	se += te;
+	s = quick_two_sum(s, se, out_precision);
+	return s;
+}
+#endif
+
 void vertex_shader(in uint instance_index, in bool is_multimesh, in uint multimesh_offset, in SceneData scene_data, in mat4 model_matrix, out vec4 screen_pos) {
 	vec4 instance_custom = vec4(0.0);
 #if defined(COLOR_USED)
 	color_interp = color_attrib;
 #endif
 
+	mat4 inv_view_matrix = scene_data.inv_view_matrix;
+
+#ifdef USE_DOUBLE_PRECISION
+	vec3 model_precision = vec3(model_matrix[0][3], model_matrix[1][3], model_matrix[2][3]);
+	model_matrix[0][3] = 0.0;
+	model_matrix[1][3] = 0.0;
+	model_matrix[2][3] = 0.0;
+	vec3 view_precision = vec3(inv_view_matrix[0][3], inv_view_matrix[1][3], inv_view_matrix[2][3]);
+	inv_view_matrix[0][3] = 0.0;
+	inv_view_matrix[1][3] = 0.0;
+	inv_view_matrix[2][3] = 0.0;
+#endif
+
 	mat3 model_normal_matrix;
 	if (bool(instances.data[instance_index].flags & INSTANCE_FLAGS_NON_UNIFORM_SCALE)) {
 		model_normal_matrix = transpose(inverse(mat3(model_matrix)));
@@ -142,11 +182,12 @@ void vertex_shader(in uint instance_index, in bool is_multimesh, in uint multime
 		model_normal_matrix = mat3(model_matrix);
 	}
 
+	mat4 matrix;
+	mat4 read_model_matrix = model_matrix;
+
 	if (is_multimesh) {
 		//multimesh, instances are for it
 
-		mat4 matrix;
-
 #ifdef USE_PARTICLE_TRAILS
 		uint trail_size = (instances.data[instance_index].flags >> INSTANCE_FLAGS_PARTICLE_TRAIL_SHIFT) & INSTANCE_FLAGS_PARTICLE_TRAIL_MASK;
 		uint stride = 3 + 1 + 1; //particles always uses this format
@@ -232,7 +273,14 @@ void vertex_shader(in uint instance_index, in bool is_multimesh, in uint multime
 #endif
 		//transpose
 		matrix = transpose(matrix);
-		model_matrix = model_matrix * matrix;
+#if !defined(USE_DOUBLE_PRECISION) || defined(SKIP_TRANSFORM_USED) || defined(VERTEX_WORLD_COORDS_USED) || defined(MODEL_MATRIX_USED)
+		// Normally we can bake the multimesh transform into the model matrix, but when using double precision
+		// we avoid baking it in so we can emulate high precision.
+		read_model_matrix = model_matrix * matrix;
+#if !defined(USE_DOUBLE_PRECISION) || defined(SKIP_TRANSFORM_USED) || defined(VERTEX_WORLD_COORDS_USED)
+		model_matrix = read_model_matrix;
+#endif // !defined(USE_DOUBLE_PRECISION) || defined(SKIP_TRANSFORM_USED) || defined(VERTEX_WORLD_COORDS_USED)
+#endif // !defined(USE_DOUBLE_PRECISION) || defined(SKIP_TRANSFORM_USED) || defined(VERTEX_WORLD_COORDS_USED) || defined(MODEL_MATRIX_USED)
 		model_normal_matrix = model_normal_matrix * mat3(matrix);
 	}
 
@@ -297,7 +345,22 @@ void vertex_shader(in uint instance_index, in bool is_multimesh, in uint multime
 // using local coordinates (default)
 #if !defined(SKIP_TRANSFORM_USED) && !defined(VERTEX_WORLD_COORDS_USED)
 
+#ifdef USE_DOUBLE_PRECISION
+	// We separate the basis from the origin becasue the basis is fine with single point precision.
+	// Then we combine the translations from the model matrix and the view matrix using emulated doubles.
+	// We add the result to the vertex and ignore the final lost precision.
+	vec3 model_origin = model_matrix[3].xyz;
+	if (is_multimesh) {
+		vertex = mat3(matrix) * vertex;
+		model_origin = double_add_vec3(model_origin, model_precision, matrix[3].xyz, vec3(0.0), model_precision);
+	}
+	vertex = mat3(model_matrix) * vertex;
+	vec3 temp_precision; // Will be ignored.
+	vertex += double_add_vec3(model_origin, model_precision, scene_data.inv_view_matrix[3].xyz, view_precision, temp_precision);
+	vertex = mat3(scene_data.view_matrix) * vertex;
+#else
 	vertex = (modelview * vec4(vertex, 1.0)).xyz;
+#endif
 #ifdef NORMAL_USED
 	normal = modelview_normal * normal;
 #endif
@@ -490,7 +553,6 @@ layout(location = 10) in flat uint instance_index_interp;
 
 //defines to keep compatibility with vertex
 
-#define model_matrix instances.data[draw_call.instance_index].transform
 #ifdef USE_MULTIVIEW
 #define projection_matrix scene_data.projection_matrix_view[ViewIndex]
 #define inv_projection_matrix scene_data.inv_projection_matrix_view[ViewIndex]
@@ -737,6 +799,17 @@ void fragment_shader(in SceneData scene_data) {
 	vec2 alpha_texture_coordinate = vec2(0.0, 0.0);
 #endif // ALPHA_ANTIALIASING_EDGE_USED
 
+	mat4 inv_view_matrix = scene_data.inv_view_matrix;
+	mat4 read_model_matrix = instances.data[instance_index].transform;
+#ifdef USE_DOUBLE_PRECISION
+	read_model_matrix[0][3] = 0.0;
+	read_model_matrix[1][3] = 0.0;
+	read_model_matrix[2][3] = 0.0;
+	inv_view_matrix[0][3] = 0.0;
+	inv_view_matrix[1][3] = 0.0;
+	inv_view_matrix[2][3] = 0.0;
+#endif
+
 	{
 #CODE : FRAGMENT
 	}